Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Talking Back -- human input and explanations to interactive AI systems

Alan Dix, Tommaso Turchi, Ben Wilson, Anna Monreale, Matt Roach

TL;DR

This work investigates turning explainable AI on its head by enabling humans to provide explanations to AI systems, with the goal of achieving more human-aligned judgments and reduced bias. It surveys multiple input modalities—ranging from feature constraints and counterfactuals to numeric importance, intermediate features, and free-text rationales—and discusses how these can be integrated into learning and explanation generation. It translates these ideas into algorithmic and architectural implications, offering concrete strategies for constraining decision trees, weighting features, and embedding explanations within interactive systems, while also addressing social and governance considerations. The paper lays groundwork for synergistic human–AI partnerships that adapt explanations to diverse domains, such as education and healthcare, and outlines concrete future work to operationalize QbB-driven explanations.

Abstract

While XAI focuses on providing AI explanations to humans, can the reverse - humans explaining their judgments to AI - foster richer, synergistic human-AI systems? This paper explores various forms of human inputs to AI and examines how human explanations can guide machine learning models toward automated judgments and explanations that align more closely with human concepts.

Talking Back -- human input and explanations to interactive AI systems

TL;DR

This work investigates turning explainable AI on its head by enabling humans to provide explanations to AI systems, with the goal of achieving more human-aligned judgments and reduced bias. It surveys multiple input modalities—ranging from feature constraints and counterfactuals to numeric importance, intermediate features, and free-text rationales—and discusses how these can be integrated into learning and explanation generation. It translates these ideas into algorithmic and architectural implications, offering concrete strategies for constraining decision trees, weighting features, and embedding explanations within interactive systems, while also addressing social and governance considerations. The paper lays groundwork for synergistic human–AI partnerships that adapt explanations to diverse domains, such as education and healthcare, and outlines concrete future work to operationalize QbB-driven explanations.

Abstract

While XAI focuses on providing AI explanations to humans, can the reverse - humans explaining their judgments to AI - foster richer, synergistic human-AI systems? This paper explores various forms of human inputs to AI and examines how human explanations can guide machine learning models toward automated judgments and explanations that align more closely with human concepts.

Paper Structure

This paper contains 16 sections, 8 figures.

Table of Contents

  1. Introduction
  2. Human inputs to AI
  3. Human explanations
  4. Feature constraints in QbB
  5. Counterfactuals in skeptical learning
  6. Feeding-back importance measures
  7. User generated intermediate features
  8. Free text -- Chatbots and LLMs
  9. Algorithmic implications
  10. Incorporating feature constraints into decision trees
  11. Incorporating feature importance measures into learning
  12. Intermediate level features and user generated vocabulary
  13. Free text explanations
  14. Architectural implications
  15. Social implications
  16. ...and 1 more sections

Figures (8)

  • Figure 1: Common communications between human user and AI.
  • Figure 2: Query-by-Browsing -- how it works dix1994query.
  • Figure 3: Some less common communication types between human user and AI.
  • Figure 4: QbB -- global feature selection.
  • Figure 5: QbB -- local feature explanations: (above) select feature(s); (below) specify local rule.
  • ...and 3 more figures